Executive Summary Cryptographic watermarking of large language model outputs—operationalized via pseudorandom tournament sampling (e.g., SynthID-Text architectures)—creates structural asymmetries across regulatory compliance, forensic attribution, and adversarial spoofing vectors. While European Union AI Act Article 50 mandates transparency, the statistical mechanics of token-distribution bias fail to differentiate between wholesale generation, light structural editing, and multi-author consensus writing. … Leggi tutto AI Watermarking Forensics: Multi-Domain Threat Analysis